Weather forecasting centers currently rely on statistical postprocessing methods to minimize forecast error. This improves skill but can lead to predictions that violate physical principles or disregard dependencies between variables, which can be problematic for downstream applications and for the trustworthiness of postprocessing models, especially when they are based on new machine learning approaches. Building on recent advances in physics-informed machine learning, we propose to achieve physical consistency in deep learning-based postprocessing models by integrating meteorological expertise in the form of analytic equations. Applied to the post-processing of surface weather in Switzerland, we find that constraining a neural network to enforce thermodynamic state equations yields physically-consistent predictions of temperature and humidity without compromising performance. Our approach is especially advantageous when data is scarce, and our findings suggest that incorporating domain expertise into postprocessing models allows to optimize weather forecast information while satisfying application-specific requirements.

Large language models (LLMs) have demonstrated strong performance in zero-shot reasoning tasks, including abductive reasoning. This is reflected in their ability to perform well on current benchmarks in this area. However, to truly test the limits of LLMs in abductive reasoning, a more challenging benchmark is needed. In this paper, we present such a benchmark, consisting of 191 long-form mystery stories, each approximately 1200 words in length and presented in the form of detective puzzles. Each puzzle includes a multiple-choice question for evaluation sourced from the "5 Minute Mystery" platform. Our results show that state-of-the-art GPT models perform significantly worse than human solvers on this benchmark, with an accuracy of 28\% compared to 47\% for humans. This indicates that there is still a significant gap in the abductive reasoning abilities of LLMs and highlights the need for further research in this area. Our work provides a challenging benchmark for future studies on reasoning in language models and contributes to a better understanding of the limits of LLMs' abilities.

Very few eXplainable AI (XAI) studies consider how users understanding of explanations might change depending on whether they know more or less about the to be explained domain (i.e., whether they differ in their expertise). Yet, expertise is a critical facet of most high stakes, human decision making (e.g., understanding how a trainee doctor differs from an experienced consultant). Accordingly, this paper reports a novel, user study (N=96) on how peoples expertise in a domain affects their understanding of post-hoc explanations by example for a deep-learning, black box classifier. The results show that peoples understanding of explanations for correct and incorrect classifications changes dramatically, on several dimensions (e.g., response times, perceptions of correctness and helpfulness), when the image-based domain considered is familiar (i.e., MNIST) as opposed to unfamiliar (i.e., Kannada MNIST). The wider implications of these new findings for XAI strategies are discussed.

While recent work on text-conditional 3D object generation has shown promising results, the state-of-the-art methods typically require multiple GPU-hours to produce a single sample. This is in stark contrast to state-of-the-art generative image models, which produce samples in a number of seconds or minutes. In this paper, we explore an alternative method for 3D object generation which produces 3D models in only 1-2 minutes on a single GPU. Our method first generates a single synthetic view using a text-to-image diffusion model, and then produces a 3D point cloud using a second diffusion model which conditions on the generated image. While our method still falls short of the state-of-the-art in terms of sample quality, it is one to two orders of magnitude faster to sample from, offering a practical trade-off for some use cases. We release our pre-trained point cloud diffusion models, as well as evaluation code and models, at https://github.com/openai/point-e.

In many high-dimensional prediction or classification tasks, complementary data on the features are available, e.g. prior biological knowledge on (epi)genetic markers. Here we consider tasks with numerical prior information that provide an insight into the importance (weight) and the direction (sign) of the feature effects, e.g. regression coefficients from previous studies. We propose an approach for integrating multiple sources of such prior information into penalised regression. If suitable co-data are available, this improves the predictive performance, as shown by simulation and application. The proposed method is implemented in the R package `transreg' (https://github.com/lcsb-bds/transreg).

The findable, accessible, interoperable, and reusable (FAIR) data principles have provided a framework for examining, evaluating, and improving how we share data with the aim of facilitating scientific discovery. Efforts have been made to generalize these principles to research software and other digital products. Artificial intelligence (AI) models -- algorithms that have been trained on data rather than explicitly programmed -- are an important target for this because of the ever-increasing pace with which AI is transforming scientific and engineering domains. In this paper, we propose a practical definition of FAIR principles for AI models and create a FAIR AI project template that promotes adherence to these principles. We demonstrate how to implement these principles using a concrete example from experimental high energy physics: a graph neural network for identifying Higgs bosons decaying to bottom quarks. We study the robustness of these FAIR AI models and their portability across hardware architectures and software frameworks, and report new insights on the interpretability of AI predictions by studying the interplay between FAIR datasets and AI models. Enabled by publishing FAIR AI models, these studies pave the way toward reliable and automated AI-driven scientific discovery.

Machine learning models have been found to learn shortcuts -- unintended decision rules that are unable to generalize -- undermining models' reliability. Previous works address this problem under the tenuous assumption that only a single shortcut exists in the training data. Real-world images are rife with multiple visual cues from background to texture. Key to advancing the reliability of vision systems is understanding whether existing methods can overcome multiple shortcuts or struggle in a Whac-A-Mole game, i.e., where mitigating one shortcut amplifies reliance on others. To address this shortcoming, we propose two benchmarks: 1) UrbanCars, a dataset with precisely controlled spurious cues, and 2) ImageNet-W, an evaluation set based on ImageNet for watermark, a shortcut we discovered affects nearly every modern vision model. Along with texture and background, ImageNet-W allows us to study multiple shortcuts emerging from training on natural images. We find computer vision models, including large foundation models -- regardless of training set, architecture, and supervision -- struggle when multiple shortcuts are present. Even methods explicitly designed to combat shortcuts struggle in a Whac-A-Mole dilemma. To tackle this challenge, we propose Last Layer Ensemble, a simple-yet-effective method to mitigate multiple shortcuts without Whac-A-Mole behavior. Our results surface multi-shortcut mitigation as an overlooked challenge critical to advancing the reliability of vision systems. The datasets and code are released: https://github.com/facebookresearch/Whac-A-Mole.git.

Many machine learning problems encode their data as a matrix with a possibly very large number of rows and columns. In several applications like neuroscience, image compression or deep reinforcement learning, the principal subspace of such a matrix provides a useful, low-dimensional representation of individual data. Here, we are interested in determining the $d$-dimensional principal subspace of a given matrix from sample entries, i.e. from small random submatrices. Although a number of sample-based methods exist for this problem (e.g. Oja's rule \citep{oja1982simplified}), these assume access to full columns of the matrix or particular matrix structure such as symmetry and cannot be combined as-is with neural networks \citep{baldi1989neural}. In this paper, we derive an algorithm that learns a principal subspace from sample entries, can be applied when the approximate subspace is represented by a neural network, and hence can be scaled to datasets with an effectively infinite number of rows and columns. Our method consists in defining a loss function whose minimizer is the desired principal subspace, and constructing a gradient estimate of this loss whose bias can be controlled. We complement our theoretical analysis with a series of experiments on synthetic matrices, the MNIST dataset \citep{lecun2010mnist} and the reinforcement learning domain PuddleWorld \citep{sutton1995generalization} demonstrating the usefulness of our approach.

Inspired by the cognitive science theory of the explicit human memory systems, we have modeled an agent with short-term, episodic, and semantic memory systems, each of which is modeled with a knowledge graph. To evaluate this system and analyze the behavior of this agent, we designed and released our own reinforcement learning agent environment, "the Room", where an agent has to learn how to encode, store, and retrieve memories to maximize its return by answering questions. We show that our deep Q-learning based agent successfully learns whether a short-term memory should be forgotten, or rather be stored in the episodic or semantic memory systems. Our experiments indicate that an agent with human-like memory systems can outperform an agent without this memory structure in the environment.

The evaluation of object detection models is usually performed by optimizing a single metric, e.g. mAP, on a fixed set of datasets, e.g. Microsoft COCO and Pascal VOC. Due to image retrieval and annotation costs, these datasets consist largely of images found on the web and do not represent many real-life domains that are being modelled in practice, e.g. satellite, microscopic and gaming, making it difficult to assert the degree of generalization learned by the model. We introduce the Roboflow-100 (RF100) consisting of 100 datasets, 7 imagery domains, 224,714 images, and 805 class labels with over 11,170 labelling hours. We derived RF100 from over 90,000 public datasets, 60 million public images that are actively being assembled and labelled by computer vision practitioners in the open on the web application Roboflow Universe. By releasing RF100, we aim to provide a semantically diverse, multi-domain benchmark of datasets to help researchers test their model's generalizability with real-life data. RF100 download and benchmark replication are available on GitHub.